I have no objection against this. Just one test case to test - create several EBR partitions (e.g. s5, s6, s7), then remove one from the middle (i.e. s6), then create two smaller in this free space (and then optionally reboot, and see what we will have after reboot).

I'll try to test this and commit with small modifications after weekend. Thanks!

It seems rt_pktsent is already unused in head/, thus its removal is reasonable. According to your calculations for offsets, this change can give some performance boost, and I'll try to measure it in the lab, but I'm not sure this will happen very soon. Also maybe is it worth to add some explicit alignment requirements to rtentry structure or some of its fields? We can use __aligned(CACHE_LINE_SIZE)

I also agree that these messages should be removed. But moving them to nd6 debugging seems wrong.

JFYI, this code will be removed when refactoring to the epoch(9) will be finished.

Do you have performance test results for already existing linux implementation?
From a quick look it seems to me there will be bottleneck regarding locking that seems can be reduced using CK and epoch. But this task can be done in future, if you plan support this code.

In D24232#535905, @melifaro wrote:

I'm going to commit it on Sunday, April 12 unless anyone has any objection.

LGTM.

how did you test this change?

Looks good to me. But I'd wait for @melifaro, there are several places where this KPI is used and I didn't check how it can affect it :)

This code, probably, was written with hope of depreciation of scope embedding into addresses.
If you plan to remove these, please leave the comment that returned address has embedded scope.
Also you will probably want to remove related XXX hack from fast forwarding code.

It seems some of these macros should be fixed to respect style(9) and use tab instead of space after #define.

Also, how did test your changes? :)
NAT64 currently is not widely used, thus changes here can break something and you will know about breakage when it will be not so easy to fix, e.g. after release.

In D23737#521593, @neel_neelc.org wrote:

Here, I also compare the destination addresses. Is this what you want?

The patch is not correct. IPv4 address can be embedded in different places depending from configuration.

I meant this thread: https://lists.freebsd.org/pipermail/freebsd-net/2019-October/054632.html

We removed this code two years ago, since it adds big penalty for routers where are tens of thousands IPv6 routes.
However, someone in the mail list recently complained that we have not a method for route expiring, probably we can rework this code to implement such feature.

Commited in rS355712.

Update the patch.

Since replay field is optional, I think you need add the check that it is not NULL.

I think there is no need in such comment.

LGTM.

The most of code that uses m_pullup() does check first, that the call is needed. Do you think that now it is considered as light enough, to do not check this?

Make address copying also conditional for consistency.
Add comment describing the goal of condition.

In D22243#486266, @lutz_donnerhacke.de wrote:

If copying the LL addr is happening every time a parent interface is added (bcopy) , why not schedule the update too unconditionally?
It will only happen on changing the parent interface anyway.

I'm agree with the change, but I have doubts that we have fixed all places in the kernel, where zone id is checked and properly initialized. Usually we keep zone id embedded in the address. Do you have some tests to check this code?